mcp-lava

mcp-lava is an MCP (Model Context Protocol) server that lets an LLM interact with a Linaro LAVA by wrapping the official lavacli command-line tool.

Instead of re-implementing LAVA's XML-RPC API, this server executes lavacli locally and returns results to the LLM in a tool-friendly way.

Why an MCP server for LAVA?

LAVA is powerful, but day-to-day workflows often involve many repetitive CLI steps:

• checking device availability and health
• locating the right job definition
• submitting jobs and monitoring states
• pulling logs and artifacts for failed jobs
• correlating failures across multiple runs
• re-running jobs to reproduce or to confirm a fix

An LLM can help with these workflows if it can reliably call the right commands and parse the outputs. mcp-lava provides that bridge:

• Reuses lavacli subcommands and output formats (many support --json / --yaml).
• Keeps server logic minimal (no duplicated API client).
• Makes automation conversational: "find the last failed run for this device type and summarize the failure", "show me the first kernel panic in the logs", etc.

Example scenarios

A few practical things you can do with an LLM + mcp-lava:

1. Triage a failure quickly

   • List recent jobs for a project
   • Detect failures, fetch logs, and summarize the suspected root cause
   • Extract the first occurrence of an error pattern (e.g. kernel panic, timeout, SSH failed)

2. Lab status overview

   • List devices and filter to offline/retired/busy
   • Show which workers are overloaded
   • Identify the most flaky device type based on recent job outcomes

3. Job submission assistance

   • Validate that a device type exists
   • Submit a job definition
   • Track the job until completion and then download artifacts

4. Re-run / reproduce / confirm a fix ("rerun" workflow)

   • Re-run the same job definition to reproduce a failure on demand
   • Re-run on a different device (or a different device type) to separate lab flakiness from real regressions
   • Re-run after changing one variable (new kernel, new rootfs, different test plan) and compare outcomes
   • Compare logs/artifacts between the original run and the rerun to spot what changed

5. Routine reporting

   • "Summarize all failed jobs in the last 24h, grouped by device type"
   • "Show the top 3 failure signatures and link to the relevant logs"

Installation

Create/activate a virtual environment, then install:

python -m venv .venv
# Activate the venv (command depends on your shell)
python -m pip install -U pip
pip install mcp-lava

Alternatively, you can also use uvx mcp-lava to run it directly without installing to your system Python.

Configuration (required)

mcp-lava authenticates to LAVA by creating/refreshing a lavacli identity at startup and then running all subsequent commands using that identity.

Set these environment variables:

• LAVA_URL - Base URL or RPC endpoint. If it doesn't end with /RPC2, the server appends it.
• LAVA_USERNAME - LAVA username.
• LAVA_AUTHENTICATION_TOKEN - LAVA token.

Optional:

• LAVA_IDENTITY - Identity name to use (default: mcp).
• LAVA_SSL_VERIFY - true/false (default: true).

A minimal way to represent variables (set them using your preferred method):

LAVA_URL=https://lava.example.com
LAVA_USERNAME=your.user
LAVA_AUTHENTICATION_TOKEN=<your-token>

Running the MCP server

This server uses the stdio transport (typical for MCP servers).

mcp-lava

Equivalent:

python -m mcp_lava.server

Tools

lava_run

Run any lavacli command.

• args must be an argv token list (a list of strings): exactly what you would type after lavacli.

Examples:

• lavacli devices list -> args=["devices", "list"]
• lavacli devices show qemu0 -> args=["devices", "show", "qemu0"]
• lavacli devices list --json -> args=["devices", "list", "--json"]

Do not pass a single space-separated string like args=["devices list"].

lava_help

Get lavacli --help, or help for a specific command group.

lava_command_tree

Introspect the installed lavacli Python package and return its command/subcommand tree.